Membership Inference Attack (MIA) aims to determine whether a specific data sample was included in the training dataset of a target model. Traditional MIA approaches rely on shadow models to mimic target model behavior, but their effectiveness diminishes for Large Language Model (LLM)- based recommendation systems due to the scale and complexity of training data. This paper introduces a novel knowledge distillation-based MIA paradigm tailored for LLM-based recommendation systems. Our method constructs a reference model via distillation, applying distinct strategies for member and non-member data to enhance discriminative capabilities. The paradigm extracts fused features (e.g., confidence, entropy, loss, and hidden layer vectors) from the reference model to train an attack model, overcoming limitations of individual features. Extensive experiments on extended datasets (Last.FM, MovieLens, Book-Crossing, Delicious) and diverse LLMs (T5, GPT -2, LLaMA3) demonstrate that our approach significantly outperforms shadow model-based MIAs and individual-feature baselines. The results show its practicality for privacy attacks in LLM-driven recommender systems.

Large Language Model (LLM) agent systems have advanced rapidly, driven by their strong generalization in zero-shot settings. To further enhance reasoning and accuracy on complex tasks, Multi-Agent Debate (MAD) has emerged as a promising framework that engages multiple LLM agents in structured debates to encourage diverse reasoning. However, triggering MAD for every query is inefficient, as it incurs substantial computational (token) cost and may even degrade accuracy by overturning correct single-agent answers. To address these limitations, we propose intelligent Multi-Agent Debate (iMAD), a token-efficient framework that selectively triggers MAD only when it is likely to be beneficial (i.e., correcting an initially wrong answer). To achieve this goal, iMAD learns generalizable model behaviors to make accurate debate decisions. Specifically, iMAD first prompts a single agent to produce a structured self-critique response, from which we extract 41 interpretable linguistic and semantic features capturing hesitation cues. Then, iMAD uses a lightweight debate-decision classifier, trained using our proposed FocusCal loss, to determine whether to trigger MAD, enabling robust debate decisions without test dataset-specific tuning. Through extensive experiments using six (visual) question answering datasets against five competitive baselines, we have shown that iMAD significantly reduces token usage (by up to 92%) while also improving final answer accuracy (by up to 13.5%).

Vision-Language Models (VLMs) excel at zero-shot inference but often degrade under test-time domain shifts. For this reason, episodic test-time adaptation strategies have recently emerged as powerful techniques for adapting VLMs to a single unlabeled image. However, existing adaptation strategies, such as test-time prompt tuning, typically require backpropagating through large encoder weights or altering core model components. In this work, we introduce Spectrum-Aware Test-Time Steering (STS), a lightweight adaptation framework that extracts a spectral subspace from the textual embeddings to define principal semantic directions and learns to steer latent representations in a spectrum-aware manner by adapting a small number of per-sample shift parameters to minimize entropy across augmented views. STS operates entirely at inference in the latent space, without backpropagation through or modification of the frozen encoders. Building on standard evaluation protocols, our comprehensive experiments demonstrate that STS largely surpasses or compares favorably against state-of-the-art test-time adaptation methods, while introducing only a handful of additional parameters and achieving inference speeds up to 8x faster with a 12x smaller memory footprint than conventional test-time prompt tuning. The code is available at https://github.com/kdafnis/STS.

To prevent misinformation and social issues arising from trustworthy-looking content generated by LLMs, it is crucial to develop efficient and reliable methods for identifying the source of texts. Previous approaches have demonstrated exceptional performance in detecting texts fully generated by LLMs. However, these methods struggle when confronting more advanced LLM output or text with adversarial multi-task machine revision, especially in the black-box setting, where the generating model is unknown. To address this challenge, grounded in the hypothesis that human writing possesses distinctive stylistic patterns, we propose Human Language Preference Detection (HLPD). HLPD employs a reward-based alignment process, Human Language Preference Optimization (HLPO), to shift the scoring model's token distribution toward human-like writing, making the model more sensitive to human writing, therefore enhancing the identification of machine-revised text. We test HLPD in an adversarial multi-task evaluation framework that leverages a five-dimensional prompt generator and multiple advanced LLMs to create diverse revision scenarios. When detecting texts revised by GPT-series models, HLPD achieves a 15.11% relative improvement in AUROC over ImBD, surpassing Fast-DetectGPT by 45.56%. When evaluated on texts generated by advanced LLMs, HLPD achieves the highest average AUROC, exceeding ImBD by 5.53% and Fast-DetectGPT by 34.14%. Code will be made available at https://github.com/dfq2021/HLPD.

Chain-of-Thought (CoT) reasoning enhances the problem-solving ability of large language models (LLMs) but leads to substantial inference overhead, limiting deployment in resource-constrained settings. This paper investigates efficient CoT transfer across models of different scales and architectures through an adaptive reasoning summarization framework. The proposed method compresses reasoning traces via semantic segmentation with importance scoring, budget-aware dynamic compression, and coherence reconstruction, preserving critical reasoning steps while significantly reducing token usage. Experiments on 7{,}501 medical examination questions across 10 specialties show up to 40% higher accuracy than truncation under the same token budgets. Evaluations on 64 model pairs from eight LLMs (1.5B-32B parameters, including DeepSeek-R1 and Qwen3) confirm strong cross-model transferability. Furthermore, a Gaussian Process-based Bayesian optimization module reduces evaluation cost by 84% and reveals a power-law relationship between model size and cross-domain robustness. These results demonstrate that reasoning summarization provides a practical path toward efficient CoT transfer, enabling advanced reasoning under tight computational constraints. Code will be released upon publication.

Academic advising is critical to student success in higher education, yet high student-to-advisor ratios limit advisors' capacity to provide timely support, particularly during peak periods. Recent advances in Large Language Models (LLMs) present opportunities to enhance the advising process. We present AdvisingWise, a multi-agent system that automates time-consuming tasks, such as information retrieval and response drafting, while preserving human oversight. AdvisingWise leverages authoritative institutional resources and adaptively prompts students about their academic backgrounds to generate reliable, personalized responses. All system responses undergo human advisor validation before delivery to students. We evaluate AdvisingWise through a mixed-methods approach: (1) expert evaluation on responses of 20 sample queries, (2) LLM-as-a-judge evaluation of the information retrieval strategy, and (3) a user study with 8 academic advisors to assess the system's practical utility. Our evaluation shows that AdvisingWise produces accurate, personalized responses. Advisors reported increasingly positive perceptions after using AdvisingWise, as their initial concerns about reliability and personalization diminished. We conclude by discussing the implications of human-AI synergy on the practice of academic advising.

As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources (e.g., model developers, users, and tools) within a single prompt context. Thus, enforcing an instruction hierarchy (IH) in LLMs, where higher-level directives override lower-priority requests, is critical for the reliability and controllability of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. Specifically, the model must first "think" about the relationship between a given user prompt and higher-priority (system) instructions before generating a response. To enable this capability via training, we construct VerIH, an instruction hierarchy dataset of constraint-following tasks with verifiable answers. This dataset comprises ~7K aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our finetuned models achieve consistent improvements on instruction following and instruction hierarchy benchmarks, achieving roughly a 20% improvement on the IHEval conflict setup. This reasoning ability also generalizes to safety-critical settings beyond the training distribution. By treating safety issues as resolving conflicts between adversarial user inputs and predefined higher-priority policies, our trained model enhances robustness against jailbreak and prompt injection attacks, providing up to a 20% reduction in attack success rate (ASR). These results demonstrate that reasoning over instruction hierarchies provides a practical path to reliable LLMs, where updates to system prompts yield controllable and robust changes in model behavior.

The rapid growth of research literature, particularly in large language models (LLMs), has made producing comprehensive and current survey papers increasingly difficult. This paper introduces autosurvey2, a multi-stage pipeline that automates survey generation through retrieval-augmented synthesis and structured evaluation. The system integrates parallel section generation, iterative refinement, and real-time retrieval of recent publications to ensure both topical completeness and factual accuracy. Quality is assessed using a multi-LLM evaluation framework that measures coverage, structure, and relevance in alignment with expert review standards. Experimental results demonstrate that autosurvey2 consistently outperforms existing retrieval-based and automated baselines, achieving higher scores in structural coherence and topical relevance while maintaining strong citation fidelity. By combining retrieval, reasoning, and automated evaluation into a unified framework, autosurvey2 provides a scalable and reproducible solution for generating long-form academic surveys and contributes a solid foundation for future research on automated scholarly writing. All code and resources are available at https://github.com/annihi1ation/auto_research.

OpenLVLM-MIA is a new benchmark that highlights fundamental challenges in evaluating membership inference attacks (MIA) against large vision-language models (LVLMs). While prior work has reported high attack success rates, our analysis suggests that these results often arise from detecting distributional bias introduced during dataset construction rather than from identifying true membership status. To address this issue, we introduce a controlled benchmark of 6{,}000 images where the distributions of member and non-member samples are carefully balanced, and ground-truth membership labels are provided across three distinct training stages. Experiments using OpenLVLM-MIA demonstrated that the performance of state-of-the-art MIA methods approached chance-level. OpenLVLM-MIA, designed to be transparent and unbiased benchmark, clarifies certain limitations of MIA research on LVLMs and provides a solid foundation for developing stronger privacy-preserving techniques.

Developing large language models (LLMs) to cooperate and compete effectively within multi-agent systems (MASs) is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce MARSHAL, an end-to-end RL framework that incentivizes Multi-Agent Reasoning through Self-play witH strAtegic LLMs in both cooperative and competitive games. MARSHAL features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, MARSHAL agent trained from Qwen3-4B develops strong strategic abilities that generalize to held-out games with up to 28.7% performance improvements. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of MASs in reasoning benchmarks. When integrated into leading MASs, our MARSHAL agent achieves significant performance gains of up to 10.0% on AIME, 6.6% on GPQA-Diamond, and 3.5% on average across all benchmarks. These results establish end-to-end RL training with self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs.